Sliced Elastic Distance for Evaluating Amplitude and Phase Differences in Precipitation Models

TL;DR

This paper introduces the sliced elastic distance, a new metric for evaluating climate models that accounts for amplitude, phase, and bias differences, especially useful for time-sensitive phenomena like monsoons.

Contribution

The paper proposes a novel sliced elastic distance metric that separately measures spatial and temporal discrepancies in climate model evaluations.

Findings

Effective assessment of CMIP6 precipitation simulations.

Quantified phase biases in Indian Summer Monsoon models.

Decomposed model errors into amplitude, timing, and bias components.

Abstract

Climate model evaluation plays a crucial role in ensuring the accuracy of climatological predictions. However, existing statistical evaluation methods often overlook time misalignment of events in a system's evolution, which can lead to a failure in identifying specific model deficiencies. This issue is particularly relevant for climate variables that involve time-sensitive events such as the monsoon season. To more comprehensively evaluate climate fields, we introduce a new vector-valued metric, the sliced elastic distance, through kernel convolution-derived slices. This metric simultaneously and separately accounts for spatial and temporal variability by decomposing the total distance between model simulations and observational data into three components: amplitude differences, timing variability, and bias (translation). We use the sliced elastic distance to assess CMIP6 precipitation simulations against observational data, evaluating amplitude and phase distances at both global and regional scales. In addition, we conduct a detailed phase analysis of the Indian Summer Monsoon to quantify timing biases in the onset and retreat of the monsoon season across the CMIP6 models.